Comparing three methods of calculating temporal ECG patterns from body surface potential maps

The objective of this study was to compare three methods for reducing temporal data from multiple-lead electrocardiograms (ECG) based on applying the Karhunen Loeve (KL) expansion. ECGs from 117 unipolar thoracic leads were recorded simultaneously for 15 seconds during normal sinus rhythm from 204 patients with underlying heart disease. The data were signal-averaged, then QRST and QRS intervals were normalized to 401 and 101 data points, respectively. Method A calculated a covariance matrix and Method B a correlation matrix from the 204 m by n data matrices (m=401 or 101, n=117). The two matrices were then factored to yield two sets of eigen functions. Method C calculated the eigen functions using an iterative approach in an attempt to approximate the results of factoring the entire 23,868 by 23,868 matrix. The matrix was dimensioned m by n, with m=102 and n=either 401 or 101. The first 102 leads were used to calculate a covariance matrix, which was then factored. The 52 eigen functions associated with the 51 largest eigen values plus the next 51 leads were included in a new data matrix and the process of calculating the covariance matrix and the new data matrix was repeated until all 23,868 leads were included. The results illustrated that while Method C was more computationally intensive, it provided the most effective method for representing the salient ECG features for both the QRST and the QRS interval based on the error assessment.